# Deep learning radiomics model of epicardial adipose tissue for predicting postoperative atrial fibrillation after lung lobectomy in lung cancer patients

## TL;DR

A deep learning model using heart fat images helps predict which lung cancer patients are at high risk for heart rhythm issues after surgery.

## Contribution

A novel deep learning radiomics model integrating clinical and image-based features for predicting postoperative atrial fibrillation in lung cancer patients.

## Key findings

- The DL radiomics model achieved AUC values of 0.890 in training, 0.876 in testing, and 0.803 in validation for predicting postoperative atrial fibrillation.
- Combining clinical features with handcrafted and DL radiomics signatures improved model performance compared to a clinical-only model.
- Machine learning algorithms and resampling techniques did not significantly enhance model performance metrics.

## Abstract

To develop and validate a deep learning (DL) radiomics model based on epicardial adipose tissue (EAT) for identifying high-risk lung cancer patients with postoperative atrial fibrillation after lung lobectomy.

A total of 1,008 patients from two centers were included. Handcrafted and DL radiomics features were extracted from the preoperative contrast-enhanced chest CT images of EAT. Clinical features and handcrafted and DL radiomics signatures were integrated to construct predictive models using the logistic regression algorithm as the baseline model. Twenty DL radiomics models were constructed through various combinations of machine learning algorithms and resampling techniques. The post hoc Nemenyi test was employed to compare the predictive performance in terms of the area under the receiver operating characteristic curve (AUC), G-mean, and F-measure.

Advanced age and male sex were identified as independent risk factors for POAF. The DL radiomics model, integrating clinical features, handcrafted radiomics signature, and DL radiomics signature, outperformed the clinical model, achieving AUC values of 0.890 (95% CI: 0.816–0.963), 0.876 (95% CI: 0.755–0.997), and 0.803 (95% CI: 0.651–0.955) in the training, testing, and validation cohorts, respectively. The results of the post hoc Nemenyi tests indicated that neither machine learning algorithms nor resampling techniques significantly improved model performance, as measured by the AUC, G-mean, or F-measure.

The DL radiomics model based on preoperative EAT images effectively identifies high-risk lung cancer patients with POAF following lung lobectomy and offers a novel tool for risk stratification.